Computer Vision News - December 2018

● A DeepLab framework was used for the mask refinement gCNN(). The backbone net is a VGG-Net. An additional skip was added, connecting intermediate pooling layers to a final output convolutional layer to enable multi-level feature fusion. ● In case of Multiple Objects -- each object is handled individually in each iteration before starting the next iteration. Overlapped regions are divided into connected pixel blobs and each blob is assigned to a label that minimizes for that blob. ● FlowNet2 used to compute optical flow. 6. Results: In the above table: TF stand for temporal fusion, MR for mask refinement. TF&MRxn means that the algorithm is performed for n iterations with both TF and MR. The table shows the results of an ablation study on the DAVIS 2017 validation set. The baseline is OSVOS. The “Boost” column calculates the performance gain for each algorithm variant. Research 8 Research Computer Vision News